Google Cloud shifts AI legal liability upward in new indemnity contracts

A compliance officer at a Swedish enterprise software vendor took a call from a Google Cloud attorney the day her team shipped a new generative AI module for retail and finance clients. The attorney pointed straight to a clause buried deep in the infrastructure contract: who pays if an AI output draws a lawsuit?

The stakes have changed for cloud risk. Major software vendors and startups are deploying generative AI products at speed, relying on foundation models operated by Google Cloud under “indemnified” terms. But with each launch, legal responsibility for what the AI produces becomes less clear. After a year of escalating suits and shifting terms, the burden moves further up the supply chain just as the market accelerates.

Google Cloud’s new indemnity commitments have brought the question of first exposure into sharp relief. Months of negotiation between SaaS vendors and platforms produced the current language, say two in-house attorneys at mid-market SaaS companies in comments to Computing Legal News on May 2026. Enterprise clients, prompted by events like the Getty/Stable Diffusion settlement and last year’s New York Times v. Microsoft filing, pressed both SaaS and infrastructure providers to define the new boundaries.

But legal risk does not evaporate for software vendors. Google’s guarantee, set out in the official terms, protects only against intellectual property claims from use of the model as prescribed—secondary consequences are expressly excluded, such as where a vendor’s own shipped product leads to a new dispute, according to Google Cloud’s Generative AI indemnity terms dated April 22, 2026.

An insurance advisor at Lloyd’s observed after reading a recent policy, “You are only indemnified up until the point your code creates a new exposure.”

Mid-market SaaS firms now face a hard line.

Google may intervene to defend if an AI-powered white-label service’s output infringes copyright, but the terms require the product's use to strictly follow Gemini’s official documentation.

Legal Tech Review in May 2026 reported contracts contested since February that show how thin this line remains.

What is not covered is the second exposure.

The largest force behind the new indemnity language has been the surge in lawsuits against users and builders of generative AI, rather than its underlying infrastructure, and only after a U.S. district court attached liability to a retail chain for AI-generated customer images did legal risk begin its move up the technology stack.

“You are only indemnified up until the point your code creates a new exposure.”

AWS and Microsoft Azure, says Marcia Henning, a cloud law specialist at London-based Henning LLP, are not yet matching Google’s stance. AWS’s latest public documentation from April 2026 omits explicit guarantees for generative APIs, and Azure’s comparable language is available only in limited cases. Henning adds, “Vendors moving quickly are effectively betting on back-channel assurances.”

Larger cloud clients are now negotiating standalone indemnity protections. Smaller businesses opt to buy separate endpoint insurance or push the risk down via their own client deals, according to a Lloyd’s of London mid-market report dated May 2026.

A public evidence trail has emerged on model endpoints.

Provider-led takedowns or usage restrictions at the API level after copyright lawsuits now render the technical/legal faultline visible to anyone watching.

Cloud risk has always been measured in uptime, latency, and API access.

The hidden risk is legal claims that manifest years out—claims that underwriters will cover only when indemnity is explicit and the chain of loss is visible. At present, few are getting paid to absorb this exposure.

For customers moving quickly, pressure is measured in weeks. Product squads launch faster than their lawyers can finish contract drafts.

For infrastructure giants, pressure unfolds over quarters, set by regulators and judges instead of release cycles.

This arrangement will buckle as soon as public documentation for Gemini or Vertex endpoints narrows coverage or expands exclusions, leaving the last-mover vendor closest to the customer holding the risk.

The legal burden moving up the stack has a precise analogue. In the early cloud era, uptime guarantees shifted from the data center to the infrastructure platform, leaving end customers scrambling only when the provider’s SLA let them down. Here too, the provider’s public stance draws the line for everyone downstream.

The next time a cloud provider limits or pulls a generative endpoint following a claim, markets will see in real time which exposures are absorbed at the infrastructure layer and which are rerouted to the application.